WaveFunctionDiffusion

Based on a mix of Wave Function Collapse (WFC) and Stable Diffusion (SD) algorithms, this repository generates a tile map (demonstrated with Starcraft:Remastered maps) from a simple text prompt (txt2img) or a given image (img2img).

It uses a dreamboothed Stable Diffusion model trained with images of tile maps, and a custom VAE model (AutoencoderTile) to encode and decode the latent variables to and from tile probabilities ("waves").

A WFC Guidance is also added to the sampling process, which pushes the generated tile map closer to the WFC transition rules. For more information about how guidance works, check out this tutorial: Fine-Tuning, Guidance and Conditioning

The model is trained with melee maps on all tilesets, which are downloaded from Battle.net, bounding.net (scmscx.com) and broodwarmaps.net.

Huggingface Model: https://huggingface.co/wdcqc/starcraft-terrain-64x64

Huggingface Spaces: https://huggingface.co/spaces/wdcqc/wfd

Run with Colab

Installation

1. Install Python 3.9 (3.8 or 3.10 should also work)
2. Install CUDA and PyTorch
3. Install requirements with pip install -r requirements.txt
4. (Optional) Install xformers (this is pretty complicated but it increases diffusion efficiency)

Local demo

Start the web UI demo:

python demo.py

Then open up the browser and navigate to the displayed URL to start diffusing!

In local demo, generated maps (.scx files) are saved in the outputs folder.

Using the 🧨diffusers pipeline

Sample code:

# Load pipeline
from wfd.wf_diffusers import WaveFunctionDiffusionPipeline
from wfd.wf_diffusers import AutoencoderTile
from wfd.scmap import find_tile_data, get_tileset_keyword

# Tilesets: ashworld, badlands, desert, ice, jungle, platform, twilight, install
tileset = "ice"

# The data files are located in wfd/scmap/tile_data/wfc
wfc_data_path = find_tile_data("wfc/{}_64x64.npz".format(tileset))

# Use CUDA (otherwise it will take 15 minutes)
device = "cuda"

tilenet = AutoencoderTile.from_pretrained(
    "wdcqc/starcraft-terrain-64x64",
    subfolder="tile_vae_{}".format(tileset)
).to(device)
pipeline = WaveFunctionDiffusionPipeline.from_pretrained(
    "wdcqc/starcraft-terrain-64x64",
    tile_vae = tilenet,
    wfc_data_path = wfc_data_path
)
pipeline.to(device)

# Double speed (only works for CUDA)
pipeline.set_precision("half")

# Generate pipeline output
# need to include the dreambooth keywords "isometric starcraft {tileset_keyword} terrain"
tileset_keyword = get_tileset_keyword(tileset)
pipeline_output = pipeline(
    "lost temple, isometric starcraft {} terrain".format(tileset_keyword),
    num_inference_steps = 50,
    guidance_scale = 3.5,
    wfc_guidance_start_step = 20,
    wfc_guidance_strength = 5,
    wfc_guidance_final_steps = 20,
    wfc_guidance_final_strength = 10,
)
image = pipeline_output.images[0]

# Display raw generated image
from IPython.display import display
display(image)

# Display generated image as tiles
wave = pipeline_output.waves[0]
tile_result = wave.argmax(axis=2)

from wfd.scmap import demo_map_image
display(demo_map_image(tile_result, wfc_data_path = wfc_data_path))

# Generate map file
from wfd.scmap import tiles_to_scx
import random, time

tiles_to_scx(
    tile_result,
    "outputs/{}_{}_{:04d}.scx".format(tileset, time.strftime("%Y%m%d_%H%M%S"), random.randint(0, 1e4)),
    wfc_data_path = wfc_data_path
)

Training with other tile maps

To train with other tile maps, first check out the dataset classes in wfd/wfd/datasets.py. You should modify the dataset classes so they output samples of your own tileset and tile images.

Starcraft has 5k+ tiles each tileset, so I did a lot of game-specific shenanigans to simplify it down. Under an easier tileset it should be fine to just randomly generate samples and convert them to images, possibly using the original WFC algorithm.

Then:

1. Train the Tile VAE (Check the config files, you can modify the hyperparameters there)

python train_tile_vae.py --image_vae <path_to_stable_diffusion_vae> --config "configs/tilenet/tilenet_sc_space_32x32.json" --output "weights/tilenet" --train_config "configs/train/train_config_32x32.json" --save_dataset_wfc <path_to_save_wfc_metadata> --data_dir <path_to_training_data> --device cuda

path_to_stable_diffusion_vae should be a downloaded vae folder on the original Stable Diffusion huggingface repository.

path_to_save_wfc_metadata is the WFC metadata that should be plugged into the pipeline, see wfc_data_path above.

2. Dreambooth the original runwayml/stable-diffusion-v1-5 (For the meaning of the options you can check out the diffusers dreambooth documentation)

accelerate launch train_dreambooth.py --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" --instance_data_dir=<path_to_training_data> --class_data_dir=<path_to_class_images> --output_dir="checkpoints" --with_prior_preservation --prior_loss_weight=1.0 --instance_prompt="an image of isometric scspace terrain" --class_prompt="an image of isometric terrain" --resolution=512 --train_batch_size=1 --gradient_accumulation_steps=2 --gradient_checkpointing --use_8bit_adam --learning_rate=5e-6 --lr_scheduler="constant" --lr_warmup_steps=0 --num_class_images=1000 --max_train_steps=10000 --checkpointing_steps 500

Note that I added custom options like --append_map_name, --brightness_fix to the dreambooth script. Check python train_dreambooth.py --help for more information.

Never-Asked Questions

Q: How do I generate maps larger than 64x64?

A: Pass in parameters width and height to the pipeline function call.

The numbers should be 8 times the map size, e.g. width=512, height=512 gives a 64x64 map, and width=768, height=768 gives 96x96.

If CUDA runs out of memory, install xformers and add pipeline.enable_attention_slicing() before calling the pipeline.

pipeline.enable_vae_slicing() and pipeline.enable_sequential_cpu_offload() may help as well for reducing memory cost.

Q: Is it necessary to train the entire VAE? Looks like the encoder can totally be skipped, only the decoding step is necessary.

A: It's not. The encoder is only slightly trained to guide the decoder, currently (for 64x64 models) only the decoder is fully trained.

Q: Why take argmax of the generated wave, instead of using it as a prior distribution for WFC? It should make the result more accurate.

A: Because it doesn't work. It frequently generates impossible wave states, making WFC unsolvable (ㅠㅠㅠ)

Q: Why choose Starcraft as the target, comparing to other tile-based games?

A: It's only because I'm more familiar with starcraft. It would also be interesting for like RPG Maker and Super Mario Maker, but I would have no idea where to get the dataset.

Q: How is this algorithm related to quantum physics?

A: It's not. I named it wave function diffusion because Wave Function Collapse did it in the first place lmao

Q: You seriously wasted over three weeks of time to create a terrain generator for a dead game???

A: Well it's all blizz's fault for trashing their games, otherwise this could have been a great project

Q: This will replace game level designing jobs and make everyone jobless!!!!🤬🤬🤬

A: While I really don't think this thing'd be able to do it, at this point we should probably stop doing jobs and wait for our AI overlords to overtake us all

Q: Can this algorithm solve sudoku?

A: You should do it yourself. It's a game.